Anthropic has introduced Haiku 5.5, a new version of its Claude Haiku series designed for high-volume, cost-sensitive applications. The model is positioned as the company’s fastest and most economical small AI system to date, targeting workloads such as summarization, data compaction, database queries, and classification. It complements Anthropic’s larger Opus 5.5 and Sonnet 5.5 models, which are used for more complex coding tasks as subagents.

Haiku 5.5 is priced at approximately 75% less than its predecessor, Haiku 4.5, according to company documentation. Alongside the launch, Anthropic announced a 50% reduction in the cost of cache reads for Sonnet 5.5, lowering the price to $0.10 per million tokens. This adjustment reduces the operational cost of Sonnet 5.5 on most agentic workflows by about 20%, as noted in internal benchmarking.

Performance data released by Anthropic shows Haiku 5.5 achieving 72.4% accuracy on the OSWorld 2.1 offline subset benchmark, up from 48.9% in Haiku 4.5. On the Humanity’s Last Exam benchmark, it scored 56.9% with tools and 18.7% without, compared to 10.2% and 0% for GPT-6 Luna in similar conditions. In agentic coding tasks, Haiku 5.5 achieved 39.2% pass rate on Terminal-Bench 4.0, outperforming Haiku 4.5’s 16.4% but falling short of Sonnet 5.5’s 70.6%. The model also scored 61.6% on Chartography, a visual reasoning test, when no tools were used.

A key feature of Haiku 5.5 is its adjustable effort setting, allowing users to balance speed and intelligence based on task requirements. Lower effort settings reduce cost, while higher settings increase computational depth. This flexibility supports use cases ranging from real-time customer support to browser automation, where latency is critical.

Pricing details indicate that Haiku 5.5 offers tiered rates based on prompt and output length. For prompts up to 100,000 tokens—which account for roughly 90% of prior Haiku usage—the cost is $0.01 per million tokens for input and $0.50 per million for output. Cache write operations are priced at $0.125 per million tokens for short inputs and $0.625 for longer ones. Output token pricing follows a similar structure, with rates increasing for extended responses.

Safety evaluations show Haiku 5.5 demonstrates improved alignment performance over Haiku 4.5, with fewer instances of misaligned behavior and reduced susceptibility to misuse. Its cybersecurity safeguards are more permissive than Haiku 4.5’s but remain more restrictive than those applied to Sonnet 5.5, blocking activities such as penetration testing while allowing defensive security tasks. Biology-related queries are permitted under controlled conditions, though requests deemed potentially harmful are restricted. Access to broader biological or cybersecurity research may require participation in Anthropic’s verification programs.

Haiku 5.5 is now available through major cloud providers, including Amazon Web Services, Google Cloud, and Microsoft Azure. On the Claude Platform, developers can deploy the model using the identifier claude-haiku-5-5. New SDK updates for Python and TypeScript introduce beta support for computer use and browser automation, aligning with Haiku 5.5’s strengths in speed and efficiency.

Subscribers to Max and Team tiers will receive monthly API credits—up to $500—for use across the Claude Platform. These credits are intended to support experimentation with AI agents, automation tools, and application development. The rollout includes expanded capabilities for developers building multi-step workflows, particularly in environments requiring rapid, repetitive processing.

Asana staff software engineer Aaron Vinh noted in internal testing that Haiku 5.5 reduced latency by over 30% compared to previous models when handling tasks such as bug triage, project setup, and portfolio risk analysis. The company described the improvement as a "noticeably snappier experience" in agent-based operations.

The release reflects a broader trend in AI development toward cost-efficient, specialized models that can be deployed at scale for routine computational tasks, while larger systems continue to handle more complex reasoning and planning workloads.